COBRA-Skills is a framework that optimizes reusable skills for LLM agents by framing skill development as budgeted sequential optimization over a dynamically evolving candidate space. It uses contextual bandit algorithms to allocate evaluation budget toward promising skill candidates while refining the skill population using execution feedback. Across six benchmarks and three target models, it matches or exceeds existing skill-optimization approaches while cutting optimization cost by 55-58%, using only 50 unique examples per benchmark.
